A Federated Model Update Method for Intelligent Gas Sensor Replacement

Abstract

Due to the limited lifespan of Micro-Electro-Mechanical Systems (MEMS), their components need to be replaced regularly. For intelligent devices such as electronic noses, updating an intelligent gas sensor system requires establishing a new classifier model for the newly inserted gas sensor probes because of the poor consistency between the signals collected by the new and original systems. The traditional method involves retraining the new model by collecting adequate data of the gas sensor array under strict laboratory conditions, which is time-consuming and resource-intensive. To simplify and expedite this process, a federated learning method called FedGSSU is proposed for gas sensor system updating. Two datasets were used to verify the effectiveness of the proposed framework. The experimental results show that FedGSSU can effectively utilize the original classifier model to obtain a new classifier model while only replacing the gas sensor array. The consistency between the new gas sensor system and the original one reaches up to 90.17%, and the test accuracy is increased by 4 percentage points compared to the traditional method. While replacing sensors with FedGSSU will reduce recognition accuracy slightly, it is more acceptable in scenarios where high accuracy is not required than re-calibrating sensors and re-training the classifier.